Context Budget Planner
Use this skill before broad, long-running, or multi-source Codex work. The goal
is to keep useful capability while reducing wasted context.
Output
Return:
- Objective and completion evidence
- Must-load context
- Defer-until-needed context
- Ignore or exclude list
- Tool and source routing
- Token-saving tactics
- Compaction-safe handoff
- Verification gates
Reference Routing
Read references/context-strategy.md for broad repo work, research-heavy
tasks, multi-agent delegation, long command outputs, compaction handoffs,
surface-routing decisions, or token/context tradeoffs.
Rules
- Preserve user constraints, approval gates, file paths, exact commands, and
named artifacts.
- Prefer current files and command outputs over memory or guesses.
- Load indexes, manifests, catalogs, and summaries before full files when that
still gives reliable evidence.
- Read a selected authoritative instruction file completely before acting on
it.
- Do not paste huge logs into the main thread. Summarize the key lines and keep
command names exact.
- Do not move secrets, tokens, auth files, session data, private memory, or
private local paths into prompts, docs, diagrams, or issue comments.
- Use subagents for bounded read-heavy work only when they reduce noise or
produce clearer evidence.
- Put durable behavior in the right surface: prompt,
AGENTS.md, skill,
plugin, MCP, hook, memory, rule, config, or script.
- Use current official docs or verified sources when external facts affect the
answer.
- Keep overlapping optional skills discoverable, but prefer the smallest
default set that covers the task without duplicate trigger noise.
Context Budget Template
Objective:
- ...
Completion evidence:
- ...
Must load now:
- ...
Defer until needed:
- ...
Ignore/exclude:
- ...
Tool/source routing:
- Local repo:
- Official docs:
- External docs:
- Browser/UI:
- Security/release:
Token-saving tactics:
- Start from indexes/manifests/catalogs before full files.
- Read only the relevant sections after symbol or text search.
- Summarize long command output instead of copying it wholesale.
- Keep exact commands, paths, versions, and failures.
Compaction handoff:
- Current decision:
- Files changed or inspected:
- Commands run:
- Known blockers:
- Next verification:
Verification gates:
- ...
When Not To Use
- Tiny single-file edits.
- Simple factual answers where no repo or external context is needed.
- Tasks where the user explicitly asks for direct execution and context is
already small.
1---2name: context-budget-planner3description: Plan token and context usage before broad Codex work. Use for repo-wide, research-heavy, multi-agent, documentation-heavy, or long-running tasks that need a context budget, source priority map, compaction handoff, and verification gates.4---56# Context Budget Planner78Use this skill before broad, long-running, or multi-source Codex work. The goal9is to keep useful capability while reducing wasted context.1011## Output1213Return:14151. Objective and completion evidence162. Must-load context173. Defer-until-needed context184. Ignore or exclude list195. Tool and source routing206. Token-saving tactics217. Compaction-safe handoff228. Verification gates2324## Reference Routing2526Read `references/context-strategy.md` for broad repo work, research-heavy27tasks, multi-agent delegation, long command outputs, compaction handoffs,28surface-routing decisions, or token/context tradeoffs.2930## Rules3132- Preserve user constraints, approval gates, file paths, exact commands, and33 named artifacts.34- Prefer current files and command outputs over memory or guesses.35- Load indexes, manifests, catalogs, and summaries before full files when that36 still gives reliable evidence.37- Read a selected authoritative instruction file completely before acting on38 it.39- Do not paste huge logs into the main thread. Summarize the key lines and keep40 command names exact.41- Do not move secrets, tokens, auth files, session data, private memory, or42 private local paths into prompts, docs, diagrams, or issue comments.43- Use subagents for bounded read-heavy work only when they reduce noise or44 produce clearer evidence.45- Put durable behavior in the right surface: prompt, `AGENTS.md`, skill,46 plugin, MCP, hook, memory, rule, config, or script.47- Use current official docs or verified sources when external facts affect the48 answer.49- Keep overlapping optional skills discoverable, but prefer the smallest50 default set that covers the task without duplicate trigger noise.5152## Context Budget Template5354```text55Objective:56- ...5758Completion evidence:59- ...6061Must load now:62- ...6364Defer until needed:65- ...6667Ignore/exclude:68- ...6970Tool/source routing:71- Local repo:72- Official docs:73- External docs:74- Browser/UI:75- Security/release:7677Token-saving tactics:78- Start from indexes/manifests/catalogs before full files.79- Read only the relevant sections after symbol or text search.80- Summarize long command output instead of copying it wholesale.81- Keep exact commands, paths, versions, and failures.8283Compaction handoff:84- Current decision:85- Files changed or inspected:86- Commands run:87- Known blockers:88- Next verification:8990Verification gates:91- ...92```9394## When Not To Use9596- Tiny single-file edits.97- Simple factual answers where no repo or external context is needed.98- Tasks where the user explicitly asks for direct execution and context is99 already small.
ucsahinn